Skip to content

Commit dc31f50

Browse files
fe3j2xge2i-3dzsufe3j2xge2i-3dzsu
committed
update html
1 parent 1bde62a commit dc31f50

File tree

1 file changed

+199
-7
lines changed

1 file changed

+199
-7
lines changed

‎SDAR/index.html‎

Lines changed: 199 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -70,6 +70,7 @@
7070
overflow: hidden;
7171
border-radius: 14px;
7272
border: 1px solid #e7edf5
73+
7374
}
7475

7576
.score-table {
@@ -179,8 +180,11 @@
179180
.col-tail.col-highlight {
180181
border-bottom: 3px solid #2e6cff !important;
181182
}
182-
183-
让竖条在圆角处也连贯（和你现有的竖条实现配合）
183+
/* 只显示左侧蓝条 */
184+
.col-highlight--left::after { display: none; }
185+
/* 只显示右侧蓝条 */
186+
.col-highlight--right::before { display: none; }
187+
184188
</style>
185189
</head>
186190

@@ -249,9 +253,9 @@ <h1 class="title is-1 publication-title">SDAR: Synergy of AutoRegression and Dif
249253

250254
<div class="is-size-5 publication-authors">
251255
<span class="author-block"><sup>1</sup>Shanghai AI Laboratory,</span>
252-
256+
253257
<span class="author-block"><sup>2</sup>University of Maryland, College Park,</span>
254-
258+
255259
<span class="author-block"><sup>3</sup>Tsinghua University</span>
256260
<br>
257261
<span class="author-block" style="font-size: 0.85em;"><sup>*</sup>Equal Contribution</span>
@@ -308,7 +312,9 @@ <h1 class="title is-1 publication-title">SDAR: Synergy of AutoRegression and Dif
308312
<h2 class="title is-3">TL;DR</h2>
309313
<div class="content has-text-justified">
310314
<p>
311-
We propose <b>SDAR</b> (<b>S</b>ynergy of <b>D</b>iffusion and <b>A</b>uto<b>R</b>egression), a new language modeling paradigm that synergizes autoregressive and masked discrete diffusion modeling strategies. The model
315+
We propose <b>SDAR</b> (<b>S</b>ynergy of <b>D</b>iffusion and <b>A</b>uto<b>R</b>egression), a new
316+
language modeling paradigm that synergizes autoregressive and masked discrete diffusion modeling
317+
strategies. The model
312318
series is continuously trained on Qwen3 and has achieved SOTA performance and speed.
313319
</p>
314320
<p>
@@ -326,10 +332,13 @@ <h2 class="title is-3">TL;DR</h2>
326332
</div>
327333
<div class="content has-text-justified">
328334
<p>
329-
We compare the performance of <b>SDAR-30B-A3B-Chat</b> and <b>Qwen3-30B-A3B-AR-SFT</b> under both dynamic and static inference settings. Additionally, we evaluate how varying the threshold in dynamic inference affects speed relative to static inference.
335+
We compare the performance of <b>SDAR-30B-A3B-Chat</b> and <b>Qwen3-30B-A3B-AR-SFT</b> under both dynamic
336+
and static inference settings. Additionally, we evaluate how varying the threshold in dynamic inference
337+
affects speed relative to static inference.
330338
</p>
331339
<ul>
332-
<li>While the static inference speed of SDAR is comparable to that of AR models, its dynamic mode achieves over <b>2×</b> speed-up over its static counterpart with almost no loss in accuracy.</li>
340+
<li>While the static inference speed of SDAR is comparable to that of AR models, its dynamic mode achieves
341+
over <b>2×</b> speed-up over its static counterpart with almost no loss in accuracy.</li>
333342
<li>The speedup effect tends to become more pronounced with increasing model size.</li>
334343
</ul>
335344
</div>
@@ -527,6 +536,189 @@ <h3 class="score-title"><span class="badge">SDAR Series</span></h3>
527536
</div>
528537
</div>
529538

539+
<!-- ========== 表 C（SDAR-sci） ========== -->
540+
<div class="score-card">
541+
<h3 class="score-title"><span class="badge">SDAR-sci vs Others</span></h3>
542+
<div class="score-wrap" style="overflow-x: auto; table-layout: auto;">
543+
<table class="score-table">
544+
<colgroup>
545+
<col style="min-width: auto;"> <!-- 第一列 Benchmark -->
546+
<col span="12" style="min-width: 130px;"> <!-- 后面7列统一 -->
547+
</colgroup>
548+
<thead>
549+
<tr>
550+
<th>Benchmark</th>
551+
<th>AR-30B-A3B-Sci</th>
552+
<th class="col-highlight col-head col-highlight--left">SDAR-30B-A3B-Sci (greedy)</th>
553+
<th class="col-highlight col-head col-highlight--right">SDAR-30B-A3B-Sci (sample)</th>
554+
<th>Intern-S1(235B-A22B)</th>
555+
<th>InternVL3-78B</th>
556+
<th>Qwen2.5-VL-72B</th>
557+
<th>DeepSeek-R1-0528</th>
558+
<th>Qwen3-235B-A22B</th>
559+
<th>Kimi-K2-Instruct</th>
560+
<th>Gemini-2.5 Pro</th>
561+
<th>o3</th>
562+
<th>Grok-4</th>
563+
</tr>
564+
</thead>
565+
<tbody>
566+
<tr>
567+
<td>MMLU_pro</td>
568+
<td>78.3</td>
569+
<td class="col-highlight col-highlight--left">80.2</td>
570+
<td class="col-highlight col-highlight--right">80.6</td>
571+
<td>83.5</td>
572+
<td>73.0</td>
573+
<td>72.1</td>
574+
<td>83.4</td>
575+
<td>82.2</td>
576+
<td>82.7</td>
577+
<td>86.0</td>
578+
<td>85.0</td>
579+
<td>85.9</td>
580+
</tr>
581+
<tr>
582+
<td>GPQA_diamond</td>
583+
<td>61.2</td>
584+
<td class="col-highlight col-highlight--left">73.7</td>
585+
<td class="col-highlight col-highlight--right">71.8</td>
586+
<td>77.3</td>
587+
<td>49.9</td>
588+
<td>49.0</td>
589+
<td>80.6</td>
590+
<td>71.1</td>
591+
<td>77.8</td>
592+
<td>83.8</td>
593+
<td>83.3</td>
594+
<td>87.5</td>
595+
</tr>
596+
<tr>
597+
<td>AIME2024</td>
598+
<td>74.9</td>
599+
<td class="col-highlight col-highlight--left">73.3</td>
600+
<td class="col-highlight col-highlight--right">76.2</td>
601+
<td>–</td>
602+
<td>–</td>
603+
<td>–</td>
604+
<td>–</td>
605+
<td>–</td>
606+
<td>–</td>
607+
<td>–</td>
608+
<td>–</td>
609+
<td>–</td>
610+
</tr>
611+
<tr>
612+
<td>AIME2025</td>
613+
<td>60.7</td>
614+
<td class="col-highlight col-highlight--left">63.3</td>
615+
<td class="col-highlight col-highlight--right">62.2</td>
616+
<td>86.0</td>
617+
<td>10.7</td>
618+
<td>10.9</td>
619+
<td>87.5</td>
620+
<td>81.5</td>
621+
<td>51.4</td>
622+
<td>83.0</td>
623+
<td>88.9</td>
624+
<td>91.7</td>
625+
</tr>
626+
<tr>
627+
<td>Livemathbench_hard</td>
628+
<td>55.4</td>
629+
<td class="col-highlight col-highlight--left">60.7</td>
630+
<td class="col-highlight col-highlight--right">57.9</td>
631+
<td>–</td>
632+
<td>–</td>
633+
<td>–</td>
634+
<td>–</td>
635+
<td>–</td>
636+
<td>–</td>
637+
<td>–</td>
638+
<td>–</td>
639+
<td>–</td>
640+
</tr>
641+
<tr>
642+
<td>LCB Code Gen V5</td>
643+
<td>51.5</td>
644+
<td class="col-highlight col-highlight--left">40.7</td>
645+
<td class="col-highlight col-highlight--right">49.1</td>
646+
<td>–</td>
647+
<td>–</td>
648+
<td>–</td>
649+
<td>–</td>
650+
<td>–</td>
651+
<td>–</td>
652+
<td>–</td>
653+
<td>–</td>
654+
<td>–</td>
655+
</tr>
656+
<tr>
657+
<td>LCB Code Gen V6</td>
658+
<td>46.3</td>
659+
<td class="col-highlight col-highlight--left">42.3</td>
660+
<td class="col-highlight col-highlight--right">51.4</td>
661+
<td>–</td>
662+
<td>–</td>
663+
<td>–</td>
664+
<td>–</td>
665+
<td>–</td>
666+
<td>–</td>
667+
<td>–</td>
668+
<td>–</td>
669+
<td>–</td>
670+
</tr>
671+
<tr>
672+
<td>ChemBench</td>
673+
<td>60.5</td>
674+
<td class="col-highlight col-highlight--left">75.1</td>
675+
<td class="col-highlight col-highlight--right">75.1</td>
676+
<td>83.4</td>
677+
<td>61.3</td>
678+
<td>61.6</td>
679+
<td>75.6</td>
680+
<td>75.8</td>
681+
<td>75.3</td>
682+
<td>82.8</td>
683+
<td>81.6</td>
684+
<td>83.3</td>
685+
</tr>
686+
<tr>
687+
<td>PHYSICS</td>
688+
<td>39.0</td>
689+
<td class="col-highlight col-highlight--left">52.9</td>
690+
<td class="col-highlight col-highlight--right">55.6</td>
691+
<td>44.0</td>
692+
<td>23.1</td>
693+
<td>15.7</td>
694+
<td>–</td>
695+
<td>–</td>
696+
<td>–</td>
697+
<td>40.0</td>
698+
<td>47.9</td>
699+
<td>42.8</td>
700+
</tr>
701+
<tr>
702+
<td>ProteinLMBench</td>
703+
<td>59.5</td>
704+
<td class="col-highlight col-tail col-highlight--left">60.7</td>
705+
<td class="col-highlight col-tail col-highlight--right">60.0</td>
706+
<td>63.1</td>
707+
<td>61.6</td>
708+
<td>61.0</td>
709+
<td>61.4</td>
710+
<td>59.8</td>
711+
<td>66.7</td>
712+
<td>62.9</td>
713+
<td>67.7</td>
714+
<td>66.2</td>
715+
</tr>
716+
</tbody>
717+
</table>
718+
</div>
719+
</div>
720+
721+
530722

531723
</div>
532724
</div>

0 commit comments

Comments
 (0)